Table 2 The details of precise training parameters, hyperparameters and hardware, etc. used for training and inference of GANs.
From: VISGAB: Virtual staining-driven GAN benchmarking for optimizing skin tissue histology
Serial | Training parameters | Precise settings |
|---|---|---|
1 | Number of training patches | 9,960 |
2 | Number of testing patches | 2,490 |
3 | Size of each training patch | 512 × 512 |
4 | Overlapping between consecutive patches | Yes (256 × 256) |
5 | Patch-level embedding | Each patch subdivided into non-overlapping 16 × 16-pixel tiles |
6 | Optimization strategy | Mixed precision arithmetic including hyperparameter tuning i.e., AdamW optimizer with cosine annealing and FP16 mixed precision |
7 | Hyperparameters | Hyperparameters determined via a systematic grid-search over learning rate, β1/β2 values, and loss‐weight combinations, followed by a targeted manual refinement to stabilize convergence and maximize staining fidelity. Final settings are: 2 × 10− 3, β1 = 0.5–0.9, β2 = 0.999 |
8 | Learning rate schedule | Decayed smoothly to zero over 200 epochs via cosine annealing |
9 | Dropout probability | 0.1 during training and weight initialization |
10 | Loss weights (GANs) | λpatchNCE = 1.5, λGAN = 1, λidentity = 1 and λcycle = 10 |
11 | Batch size | 8 |
12 | Hardware | High-end NVIDIA A100 GPU (80GB VRAM) |
13 | Number of epoch | 200 |
14 | Mode collapse analysis | Yes. After every 50 epochs |
